SK Hynix has backed Etched, a US AI inference startup building rack-scale systems with proprietary Low Voltage Inference (LVI) and Cluster Scale Memory (CSM) technologies.
Etched solves a critical industry bottleneck: traditional GPUs offer excess compute power yet insufficient memory for large-scale AI inference, leading to cumbersome multi-tier KV caching frameworks. Its tailored inference chips reduce power usage and outperform conventional GPU HBM in memory bandwidth and capacity, mitigating KV caching-related latency and design complexity.
This investment comes after Israeli startup Majestic Labs exited stealth mode, with the firm also building custom-processor inference servers with large-scale shared memory pools.
Founded in 2023 in San Jose by three Harvard undergraduates who dropped out, Etched’s core team comprises CEO Gavin Uberti, President Robert Wachen and Chris Zhu. The team obtained a two-year $250,000 Thiel Fellowship to fuel its early development.
The team built customized LVI prefill processing and CSM shared memory to boost inference efficiency. Its full-stack solution, integrating in-house chips, packaging, PCBs and interconnection modules, enables large-scale commercial inference deployment.
Zhu pointed out mainstream AI chips suffer severe thermal throttling under high load: elevated power consumption lowers clock speeds, cutting real inference throughput to under half its theoretical peak. Etched’s chips run arithmetic units at half the typical industry voltage, significantly boosting compute density.
He also addressed key hardware tradeoffs: HBM-equipped AI chips lag behind SRAM in decoding speed due to memory and interconnection limits, while pure SRAM chips compromise compute density and capacity for faster access, reducing overall throughput.
This hinders large MoE model operation, as cross-layer token transmission across multiple memory layers builds up latency. Etched’s architecture removes redundant memory layers to resolve the issue.
Etched’s CSM architecture creates a unified low-latency shared memory pool for scalable hardware, leveraging proprietary high-speed interconnection to accelerate cross-chip memory access.
Its hybrid HBM/SRAM design balances memory capacity and cross-chip latency, delivering high-throughput interactive inference. CSM avoids the cost, stability, thermal and performance flaws of pure SRAM, 3D DRAM and optical chip solutions.
This innovation decouples HBM stacks from GPU architectural constraints, consolidates dispersed memory resources, and delivers SRAM-level access speed via advanced interconnection technology.
Etched’s rack system features a high-interconnection layout, with 16 dedicated cables connecting rear ports to eight core hardware units.
The system supports mainstream LLMs, MoE models and non-transformer architectures like Mamba, with the company focusing on full-system R&D and commercial sales.
Etched Funding History
2022 – Company founded
2023 – $5.4M seed round
2024 – $120M Series A
2025 – $500M Series B
2026 – $300M Series C
Total funding to date: $925.4M
Sequoia Capital led Etched’s 2026 Series C round, with participation from Andreessen Horowitz, Jane Street and SK Hynix. The round values the firm at $10.3 billion, double its 2024 Series B valuation.
Its consistent annual fundraising pace mirrors ongoing product iteration, backed by robust customer demand and solid investor confidence.
Wachen said the new capital will speed up inference cluster mass production. The firm has opened a 10-MW production site in Milpitas to enhance prototyping and manufacturing efficiency.
Chief Architect Saptadeep Pal disclosed Etched has over 400 employees with booked customer orders exceeding $1 billion. Its Taiwan R&D and manufacturing base enables 24-hour operation, optimizing supplier cooperation and product testing to support rapid scaling.
Its core clients cover large enterprises, emerging cloud providers and global hyperscalers offering inference-as-a-service solutions.
Comment
If Etched and Majestic Labs’ new inference racks move beyond traditional KV caching, they will reshape the enterprise storage market. These systems demand high-speed, low-latency storage, creating collaboration opportunities for vendors like Dell, HPE, NetApp and VAST Data to pair storage arrays with next-gen inference servers for optimal performance.
Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!
Your message must be between 20-3,000 characters!
English
